Combining Agents and Ontologies for Building an Intelligent

Tutoring System

Panagiotis Stamatis

, Ioannis Panagiotopoulos

, Christos Goumopoulos

1,2

and Achilles Kameas

Educational Content, Methodology and Technology Laboratory (e-CoMeT Lab), Hellenic Open University, Patras, Greece

Information and Communication Systems Engineering Department, Aegean University, Mytilene, Lesvos, Greece

Keywords: Agent-based Systems, Intelligent Tutoring Systems, Ontologies, Personalized Learning, Distance Learning,

Dynamic Courseware Generation.

Abstract: In this paper an approach for building an intelligent tutoring system is presented, based on a multi-agent

architecture and combined with ontologies for knowledge representation. The system developed is focused

on a bottom up, reactive generation of an active sequence of knowledge units regarding a set of adjustable,

high level learning goals. The learning process begins with a set of simple learning goals that require a few

learning objects and as the educational process proceeds, the student has to achieve higher learning

outcomes that combine other low level outcomes which have been already achieved. The system is able to

adapt to student’s learning profile and progress by applying proper learning tactics to prioritize through a

weight calculation scheme the sequence of the learning outcomes to achieve. The main components of the

system consisting of ontological models of the learner and the subject under study, gateway agents and tutor

agents with their core modules (learning space management and learning tactics control) are explained and a

detailed description of their interaction is given in the context of an example application. Finally, the

advantages of the proposed approach are laid out, especially in the setting of a distance learning education

system.

1 INTRODUCTION

The term Intelligent Tutoring Systems (ITSs) refers

to complex tutoring systems that can be adapted to

the needs, characteristics and learning progress of

the individual learner (Polson & Richardson, 1988).

These systems exploit a large amount of educational

knowledge and usually they employ pedagogical

methodologies. Especially in the case of agent-based

architectures, the interaction between the different

components of the ITS is achieved through the

communication of the intelligent agents assigned to

each component. A typical architecture of an ITS

consists of four models: (a) the domain model which

contains all the knowledge and problem-solving

strategies to be learned, (b) the student model which

is an overlay of the domain model; it is the core

component of an ITS and stores all the data about

student’s characteristics and progress, (c) the

tutoring (or pedagogical) model which contains all

the information about the various pedagogical

decisions and methodologies and (d) the user

interface (UI) which enables the communication

between the user and the system (Nkambou et al.,

2010).

In this paper we introduce a pilot educational

system that enhances personalized learning of

students in the context of selected courses. We

propose an agent-based intelligent tutoring system,

able to adapt to student’s characteristics by

employing learning tactics based on the student’s

learning profile and progress. More specifically, our

proposed multi-agent system architecture employs a

set of homogenous student-dedicated tutor agents

for each course. Each agent builds an internal

learning model based on the domain and available

resource semantic representation while during the

educational process the agent updates the model

based either on the student’s learning profile and

interaction or by accessing the student’s progress

with respect to a given group. The tutoring system is

not domain specific while the pedagogical module is

versatile, allowing tutors to experiment on different

learning tactics in order to engineer more domain-

specific or student profile-oriented agents. Finally,

the system self organizes student groups based on

overall group progress indicators and without any

Panagiotis S., Panagiotopoulos I., Goumopoulos C. and Kameas A..

Combining Agents and Ontologies for Building an Intelligent Tutoring System.

DOI: 10.5220/0005422900150024

In Proceedings of the 7th International Conference on Computer Supported Education (CSEDU-2015), pages 15-24

ISBN: 978-989-758-107-6

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

tutor interference. To the best of our knowledge, this

is the first indirect approach towards self-organized

learning.

In the context of our work we have chosen to

model the main components of the proposed system

through ontologies. Ontologies have been widely

used especially in the field of education and

specifically in tutoring systems for three main

reasons: (i) to support the formal representation of

abstract concepts and the relations between them in

a reusable and extendable way, (ii) to allow the

extraction of new knowledge by applying inference

mechanisms and (iii) to provide rich semantics for

humans to work with and the formalism for

computers to perform mechanical processing.

Furthermore, ontologies facilitate the reuse and the

integration of services and thus e-learning systems

are able to provide better applications (Peña &

Sossa, 2010).

Agent technology is a well-accepted approach to

address the challenges of technology enhanced

learning. In our case, by using intelligent agents in a

distance education system it is possible to obtain

adaptivity to each individual student’s learning

capabilities, particularities and learning progress.

The proposed tutoring system follows a 3-tier

architectural style. In the presentation tier users

connect to the system through a web interface; the

logic tier consists of a multi-agent system; agents

connect with a semantic repository in order to access

the domain related reusable learning objects and

student profiles (data tier). The multi agent system is

implemented using the Java Agent Development

Framework (JADE), a middleware for the

development and execution of peer to peer

applications following the agent-based development

paradigm (Bellifemine et al., 2003).

The system has been designed in the context of

the Hellenic Open University (HOU). HOU has a

mission to offer university level education using

distance learning methodology and to develop the

appropriate material and teaching methods.

Currently, HOU offers 31 undergraduate and

postgraduate Study Programs with a total of

approximately 30,000 students, coached by 1,700

tutors in 1,550 groups (20 students per group on

average). Students of the HOU usually live in

disparate locations all over the country. Besides

being students they usually have families and

working obligations so they have pressing time

constraints for studying. Given the special

characteristics of an adult distance learning

education system, the provision of tools, such as the

one presented here, that can facilitate the learning

process and enhance the learning experience are of

great importance.

The rest of the paper is structured as follows.

Section 2 provides related work on agent-based e-

learning systems. In the following section we

elaborate on the ontological models we have

implemented in order to represent the learners and

generally the knowledge of the domain to be taught.

Section 4 presents our agent-based ITS architecture

focusing on the tutor agent organization and logic. A

detailed system usage example is provided in order

to demonstrate system’s functionality. The next

section contains a discussion of the developed

system and provides future directions of work for its

improvement. Finally our conclusions are given.

2 RELATED WORK

Multi agent system (MAS) is a technology where its

application came into existence during 1980’s. A

number of e-learning systems use the multi agent

scheme to create sophisticated environments in order

to achieve maximum effectiveness in learning by

implementing different technologies and using

different methodologies (Bokhari & Ahmad, 2014).

An example in the domain of multi-agent e-learning

systems is (Ali et al., 2010) where the authors

present a multi agent approach for designing an e-

learning system architecture. The proposed

architecture consists of four tier layers, namely

Interface layer, Middle layer, Database Controller

layer and Database layer. The middle layer is based

on MAS and supports any information

communication, login, logout and new user sessions

creation. Another example of a multi agent system

that exploits ontologies for describing the

educational material as well as the learners and their

learning styles is presented in (Dung & Florea,

2011). The authors here present an architecture to

support a multi-agent e-learning system, where

intelligent agents are capable of providing

personalized assistance according to learner’s

learning style and knowledge level. A study by

(Hammami et al., 2009) describes an architecture

composed of four multi-agent system levels

interacting with each other using intelligent

blackboard agents; blackboard agents facilitate the

cooperation and coordination among interacting

agents. Each level consists of different agents

specialized on interfacing, authoring and learning

aspects depending on the human user role. The

system is connected to a number of databases

CSEDU2015-7thInternationalConferenceonComputerSupportedEducation

modelling the student profile, the learning process,

the learning domain, teaching material and practices.

The authors in (Acampora et al., 2010) apply a

Memetic Computing methodology into a

hierarchical multicore multi-agent system while

formalizing memetic agents’ exploration of

taxonomic knowledge as an optimization problem in

order to compute personalized learning experiences.

Their approach includes building a set of knowledge

highways whose paths connect information sources,

earner’s requirements and cross feasible learning

contents. Memetic agents explore the available

learning knowledge taking into account hardware

details of the available computing resources. The

domain model employs a semantic representation of

the educational domain including a set of teaching

preferences; the learning presentation generation

algorithm uses a predefined learning path of

concepts to be covered and generating the best

sequence of learning activities to best satisfy the

concept path. In (Yaghmaie & Bahreininejad, 2011)

the authors suggest a framework for building an

adaptive Learning Management System (LMS). The

proposed architecture is based upon multi-agent

systems and uses both Sharable Content Object

Reference Model (SCORM) 2004 and Semantic

Web ontology for learning content storage,

sequencing and adaptation. Moreover, they provide

a way to adapt course topics according to learners’

experiences whose learning style is similar to the

current learner.

3 ONTOLOGICAL MODELS

Ontologies are used for modelling the learners and

the knowledge of the learning domain. In order to

develop the ontologies we have followed a widely-

adopted methodology, described in (Noy &

McGuiness, 2001) and for their representation we

adopted the Web Ontology Language (OWL)

(McGuinness & van Harmelen, 2004). The

implementation process was done by the Protégé

tool which is the most widely used and offers a

complete development environment. Below we give

a more detailed description of the ontological

models we have implemented.

3.1 Learner Model

In order to implement the learner model we were

based, on the one hand, on student modelling

standards (Smythe et al., 2001, LTSC Learner

Model Working Group of the IEEE 2000, 2000) and,

on the other, on empirical studies that were

conducted by social scientists among students of

HOU. The proposed learner model is thus a

combination of stereotype and overlay techniques. A

fully stereotype-based profile, as the information

derived from student’s descriptions or questionnaires

is not accurate for every knowledge domain and the

system would adapt to student’s needs very slowly.

Dynamic attributes related to the learning process

are represented with an overlay model. From the

empirical studies we extracted information about the

dimensions/characteristics of the learner profile that

could affect his/her academic performance. A few

examples of these dimensions are: the learning style,

previous experience, reasons for education,

computer literacy, etc. The values for these attributes

(i.e. stereotypes) are used for the initialization of the

learner’s profile and then, after the initialization

phase the profile is dynamically modified as the

overlay model is updated through the interaction of

the user with the system.

The proposed student model is partially based on

the standards we mentioned above, but they also

have limitations as they reflect different perspectives

on the attributes of a learner (e.g. classic CV notion

based or student’s performance as the most

important information). On the other hand, as

resulted from the study of other similar student

models, there is no approach that satisfies all the

attributes of an adult learner within a distance

learning environment.

The learner’s model: (1) is a dynamic model that

can change over time as the system collects

information about the user, (2) is a long-term model

that keeps generalized information about the user

and not only for the current interaction with the

system and (3) combines “active” and “passive” user

modelling techniques, i.e. in the beginning user

provides direct information about him/her and then

the system collects data indirectly. The proposed

ontology defines the following four upper level

classes: (a) Student which represents any student, (b)

StudentCourseInformation which holds information

about learner’s academic performance during the

entire educational process, (c)

StudentCurrentActivity which captures learner’s

activity for the current academic year and (d)

StudentPersonalInformation which is the most

compact class of the proposed ontology,

representing not only learner’s static data, such as

demographics, but also more complex characteristics

that concern his/her interaction with the system. A

detailed description of the ontology is provided in

(Panagiotopoulos et al., 2013).

CombiningAgentsandOntologiesforBuildinganIntelligentTutoringSystem

3.2 Learning Objects and Outcomes

In order to represent the domain knowledge we have

used the notion of Learning Objects. A learning

object (LO) is defined as “a self-contained and

independent unit of digital educational content,

which is associated with one or more learning

objectives and it has as primary aim the ability of

reuse in different educational contexts”

(Nikolopoulos et al., 2012b). The ontological

representation and description of LOs has been

based on the metadata schema proposed in

(Nikolopoulos et al., 2012a).

The system also needs to keep a record of the

learner’s performance. To achieve that, we have

used the notion of Learning Outcomes. According to

the Bologna project (Bologna Working Group,

2005) a learning outcome is a statement of “what a

learner is expected to know, understand and be able

to demonstrate after completion of a learning

process (a lecture, a module or an entire program),

which are defined in terms of knowledge, skills and

competence”. For the classification of the learning

outcomes in different level skills we have applied

the Revised Bloom’s taxonomy (Krathwohl &

Anderson, 2001), as it is the most widely used. The

detailed description of the ontological representation

for the learning outcomes is given in (Kalou et al.,

2012).

It is worth noting here that the process for the

cognitive domain representation from which we

construct the corresponding learning objects and

also the definition of the learning outcomes for the

different cognitive domains is realized within a well-

defined and applied collaborative methodology

between domain experts (tutors) and knowledge

engineers, described in (Panagiotopoulos et al.,

2012, Nikolopoulos et al., 2013).

4 AGENT-BASED PLATFORM

4.1 System Architecture Overview

In this section, we provide an overview of the ITS

prototype, called APLe (Agents for Personalized

Learning), whose system architecture is depicted in

Figure 1. Students interact with the platform through

a web based interface; a servlet keeps track of all

available Gateway Agents (GA) and acts as

dispatcher in order to route each student

request/action to the proper GA, based on user and

selected course information. GAs are special

purpose agents that interface between agents of a

remote agent environment and the servlet. GAs

maintain and utilize student to Tutor Agent (TA)

mappings in order to transfer request/action

messages between students and corresponding

agents (inside their particular agent environment). In

Figure 1 two such environments are depicted

representing two different courses (e.g. Structured

Programming in C and Software Engineering). For

different students attending this course the system

will spawn different TAs; with red color we depict a

student UI-TA mapping example.

In case a GA has no mapping for a particular

user (e.g. on user login), it creates a TA. Then,

requests/actions are transformed into specific data

structures (we refer to these as Blackboard Beans or

simply beans) which encapsulate request/action

specific information such as session id, user id,

course, action type and action data. GAs are

triggered either by incoming beans or FIPA ACL

(Foundation for Intelligent Physical Agents (FIPA),

2002) compliant messages. If the agent receives a

bean, the agent translates the data into an ACL

message which is transmitted to the corresponding

agent. On the opposite, when the agent receives an

ACL message from a TA then the agent a) finds the

corresponding bean, b) attaches the agent

action/response data to the bean and c) sends the

bean back to the servlet. TAs and GAs are grouped

inside a set of agent environments (JADE

containers) which can be distributed in different

physical places of the service provider infrastructure.

Each agent environment contains at least one GA

which is created on system startup. On the other

hand, TAs are generated and terminated

dynamically.

Figure 1: System Architecture.

Although this approach introduces some extra

communication overhead (there exist user

CSEDU2015-7thInternationalConferenceonComputerSupportedEducation

actions/requests that do not affect the learning

process and thus could be handled in a central

fashion), we argue on transferring the information

load to a dedicated tutor agent for scalability, error

tolerance and security reasons: No matter of system

traffic, the servlet and GAs do not consume

resources on data processing and communication

with the repositories. If for some reason some TAs

fail, these agents are terminated while the system

continues to operate for other connected users.

Moreover, in order to deal better with any occurring

bottleneck delay, multiple gateway agents may exist

in each agent environment; finally, JADE agent

mobility can be applied to allocate agents in more

remote containers in respect of a particular grouping

policy, e.g. container traffic. The current

implementation employs a simple agent grouping

policy: TA grouping/placement is applied in regard

of the course that is currently attended while each

group (environment) contains exactly one GA.

The data tier consists of a semantic repository

and a content repository. The content repository is a

data storage facility for the available educational

content that is available for presentation. The

semantic repository contains semantic representation

and instances for the students (student profile and

action log), Learning Objects, Learning Outcomes

and finally the Domain concepts for each available

course. As a semantic repository, we use OWLIM-

Lite, a high-performance semantic repository

implemented in Java and packaged as a Storage and

Inference Layer (SAIL) for the Sesame openRDF

framework (Bishop et al., 2011). Each TA is able to

interact with the semantic repository through the

respective OWLIM-Lite by using a set of predefined

SPARQL query and update patterns. Each pattern is

defined in respect of the structural and relational

properties of the ontologies used for the semantic

representation.

4.2 Tutor Agent

Each Tutor Agent is allocated to a student that

attends a particular course at the current point of

time. The Tutor Agent (TA) architecture consists of

two modules, as depicted in Figure 2: Learning

Space Management (LSM), reflecting the internal

representation and Learning Tactic Control (LTC),

reflecting the learning tactic decision. Sesame refers

to persistence/metadata, whereas LP Updates refer to

the user feedback.

LSM is triggered either when the agent receives

a learning request/action message either on an LTC-

generated internal event. LTC may be triggered in a

Figure 2: Tutor Agent architectural design.

three way fashion: periodically, on message arrival

or on LSM-generated internal event.

Taking under consideration the agent

architecture, courseware generation is affected not

only as a result to a student’s learning events but

also in accordance to the relative distance between

the student's indicators and the group's indicators.

According to this, the TA reacts to direct and

indirect stimuli:

• Direct stimuli refer to student learning events

(or student learning event chains) which affect

directly the node state set of the learning space.

The agent reacts to the state change by guiding

the student towards a set of (most fitted)

learning outcomes. The set is produced after

applying the dominant learning tactic to the

personal learning space.

• Indirect stimuli refer to indirect interaction

between students through specific learning

indicators which are estimated globally during

the progress of the group of agents for a

particular course. The agent reacts (through

LTC) to indicator changes by switching the

(most fitted) learning tactic.

Learning Space Management

This module creates and operates upon a complex

graph structure which represents the personal

learning map of a student. The learning space is

modeled as a 3-color graph using a variety of links

based on explicit/implicit properties which are

extracted from the combined educational ontology.

Each node is described by its type, current state and

current value. The state of each node is determined

according to the state of its connections. For that

purpose, a set of well defined, non-recursive, non-

overlapping transition rules are applied on

initialization or after an educational update

(incoming message).

CombiningAgentsandOntologiesforBuildinganIntelligentTutoringSystem

The module is able to parse the graph and extract

filtered information based on a particular learning

tactic and the student profile. The outcome may be

either high level (objective) recommendation (based

on the learning tactic) or low level (learning object)

recommendation based on a student-selected

objective and the student profile. Finally, the module

updates the ontology repository (the student profile

in particular) upon each update.

Learning Tactics Control

In educational context, a learning tactic is the way a

student is attempting to learn something (Popham,

2011). We define an agent learning tactic as the way

an agent selects the next Learning Outcome for a

student to learn (to persist). More specifically, a

learning tactic is a set of connection types and

corresponding weights that apply to each node based

on the status of its very local neighbourhood

(directly connected nodes).

The Learning Tactic Control (LTC) is a reactive

selection mechanism which uses global and local

(internal) indicator updates in order to select one

learning tactic to apply to LSM. Each time the LTC

is triggered, a series of queries is sent to the

ontology repository concerning some quantitative

data about the class (or group of students). Next,

using a formula that is solely based on indicator

data, the available learning tactics are hierarchically

checked to take the control of the LSM. The

hierarchical winner-takes-all mechanism is based on

Brook’s subsumption architecture: when triggered,

top level behaviours (in our case, learning tactics)

suppress lower level behaviours from triggering.

When the dominant learning tactic switches, LTC

triggers an internal event which forces LSM to

comply with the dominant learning tactic by

resetting the connection weights of the learning

graph according to the new learning tactic. The

result of this action is a rebalanced learning graph.

4.3 System Usage Example

In this section we describe the execution phase of

the tutoring system each time a student connects to

the system, based on the following assumptions: a

tutor of the “Structured programming in C” course,

has created an initial learning plan of a single

learning goal (objective): “PA_PLI10_46_”. For that

purpose, we employ the educational ontology

discussed in Section 3, consisting of Learning

Outcomes, Learning Objects, the Bloom taxonomy

schema, an educational domain schema and finally a

student profile schema. More particularly, the

combined ontology contains 124 classes, 55 object

property types, 45 data property types and 737

individuals, including 109 Learning Outcomes, 128

Learning Objects and 208 C programming specific

Learning Concepts. A Learning Outcome has a

natural language description, an assigned Bloom

level, a number of connections with relative

Concepts and a number of connections with relative

Learning Objects. For example, “PA_PLI10_46_”

refers to “combining operators and operands in a

program to form expressions”, it is related with C

concepts like “operator”, “operand” and

“expression” and it is satisfied with Learning Object

“MA_PLI_25”. The latter is titled “Common

mistakes on using operators” related with Concepts

“operator” and “operand”, it refers to a document-

formatted example (resource type). The “operator”

concept is connected with parent concepts like

“expression” and a number of child concepts like

“Logical Operator”, “Bitwise Operator” and

“Numeric Operator”.

Also, we consider a set of two learning tactics: a)

a rapid-advance strategy which focuses on selecting

learning objectives towards higher goals as far as at

least one sub-objective is fulfilled and b) a greedy

strategy which focuses on achieving all sub-

objectives before moving toward a higher goal. The

first learning tactic is triggered by using two

indicator sets: student versus mean class quantity of

learning goals achieved multiplied by the mean class

versus student self-evaluation score. The latter tactic

is triggered by using a formula of two indicators:

student quantity versus mean class quantity for the

successful learning objectives. Also, the rapid

advance learning tactic suppresses the greedy tactic.

When a student connects to the course, a Tutor

Agent spawns inside the multi agent “Structured

Programming in C” container; next, the agent

initializes the learning space using the initial set of

objectives according to the learning plan. Next, the

graph is populated and connected recursively with

learning objectives, objects and concepts according

to a breadth first strategy using a defined set of

connection types. Currently, the exploited

connection types are five: “satisfies” between a

Learning Object and an Objective; “subject”

between a Learning Object/Objective and a Concept;

“hasBloomLevel” for Learning Objectives and the

Bloom level; finally “

hasParent” / ”hasChild”

between Concepts. The learning space generation

algorithm is set to expand uniformly all possible

connection chains with maximum Concept distance

2 from each initial learning objective (#46 in our

CSEDU2015-7thInternationalConferenceonComputerSupportedEducation

case). Using the initial set of the learning plan, the

generated learning space graph involves 46 Learning

Outcome, 52 Learning Object and 93 C

programming Concept nodes. Next, the learning

space synchronizes according to the student relevant

data from the student log. In our scenario, the course

has just started so there is no relevant data in the

student log. At this point, the student is able to use

the recommender. A graphical representation of the

learning space is depicted in Figure 3. This graph

represents a fraction of the learning map that is built

to support the goal of learning the semantics of C

operators. Learning Objects are not shown for

clarification reasons.

When the student selects the recommendation

button, the event is passed to the tutor agent who

calculates and returns back a list of the most valued

objectives of the learning space, according to the

dominant learning tactic. Each learning tactic applies

to each Learning Objective (node) of the learning

space as follows:

• rapid-advance: each node x estimates its score

based on the formula:   1/1   ∗

#

,

/#



∗#





#

,

/#



where  is

the distance of a node from the closest learning

goal, #

/

is the number of connected

incoming/outgoing nodes and

#

/,

is the number of finished

incoming/outgoing nodes. If there are no

incoming nodes, #

,







1. If there are no outgoing nodes,

#



#

,

/

#



1

• greedy: each node x estimates its score based on

the formula:

  #

,

#

,

#



, where #

,

is the

number of (incoming) nodes that are not

finished. If there are no incoming nodes,

  1.

To better understand how a learning tactic

affects the recommendation, consider Learning

Objective nodes 16, 19, 36, 27, 13 and 47: assuming

there are no (visited/finished) nodes, the former

learning tactic will estimate values ½, 0, 0, ⅓, ⅓ and

0 respectively. According to this, the rapid-advance

tactic will recommend the sequence 16, 27 and 13.

The latter tactic will estimate 1, -1, -1, 1, 1 and -

1 respectively, leading to a random recommendation

sequence for 16, 27 and 13, since all nodes have the

same weight. If we assume nodes 16 and 21 as

Figure 3: Simplified representation of the learning space in the context of the example course.

CombiningAgentsandOntologiesforBuildinganIntelligentTutoringSystem

visited, the values for 19 and 27 are not affected

(neighbours are unchanged). Nodes 36, 13 and 47

are affected, giving estimations ½, ⅙, ½ for the

former and 1, 0, 1 for the latter learning tactic. Thus,

the rapid-advance tactic will recommend the

sequence 36, 27 and 13, whereas the greedy tactic a

random sequence between 36, 27 and 47.

When the student selects an objective to attain,

the selection is passed to the agent who calculates

and returns back a list of the existing learning

objects with respect of the selected objective and the

student preferences, located in the student profile

(Figure 4) (the language used in the user interface is

currently Greek). For example, if the student prefers

visual content and the objective concerns the topic

“recursive functions”, a video learning object will be

selected, if available explaining this topic. It is noted

that the agent sorts instead of excluding learning

format/types. Thus, the student is able to select a

learning object of his/her choice.

Figure 4: Screenshot of the APLe System: Choosing

learning objects for a specific learning outcome.

Finally, when the student finishes the study of a

particular learning object, the student has to self-

evaluate his/her understanding on the learning

object. This action triggers the agent to update the

learning space and the student data repository. Also,

the agent updates its indicator data for the class.

5 DISCUSSION

The advantage of the agent-based platform derives

from the fact that the tutor agents can provide

recommendation on a sequence of learning

outcomes that most fit the student profile, according

to the properties of the learning objects. On the other

hand, the use of ontological models for representing

and storing the information regarding the learner and

the learning material enhances reusability of this

information and promotes interoperability with

third-party systems.

The research work described in this paper acts as

a proof of concept and there are still challenges that

are related with the particular approach for

engineering an ITS. Further study and experiments

will follow in order to verify well fitted agent

configurations (learning tactics and LTC) under

different domains. A research direction is to identify

the (bipartite) application of student stereotyping

into dynamic reconfigurations and vice versa.

According to these, the next step is to evaluate

the proposed system with real data from the students

of the Hellenic Open University (HOU). In order to

evaluate the proposed agent-based tutoring system,

our approach involves evaluating the system through

user's experiences to find out the usability and

impact of the ITS, finding learning rates and

achievements level.

We already have prepared a number of learning

objects with different metadata and for various

knowledge domains. These objects have different

file formats (video, document, presentation, etc.) and

different resource type (activity, exercise, self-

assessment, etc.). The criteria for the evaluation of

the APLe, have emerged from the study of different

system evaluation methodologies (such as TAM2)

and are represented through scored questionnaires

that will be given to the students of the HOU. For

example, there are questions about the usability of

the system, the interface and knowledge acquisition.

In HOU, we are developing (in a collaborative

effort among ontology experts and course tutors)

educational ontologies for the majority of the 600

courses we offer. Our aim is to gradually introduce

these ontologies to the platform and deploy the

respective agents for each course. Currently, about

40 courses are in the “pipe-line”. The platform will

eventually become a component of the HOU

educational portal, which will offer a personalized

learning environment to our students. In its first

deployment, the course ontologies will be

independent, thus a different instantiation of the

platform per course is planned. This approach will

also help us sidestep scalability issues and allow us

measure system’s performance, so as to plan the

next deployment phase.

CSEDU2015-7thInternationalConferenceonComputerSupportedEducation

6 CONCLUSIONS

This paper presented an integrated intelligent

tutoring system in order to support distance learning,

especially for adult learners. The proposed

architecture is based on a multi-agent system which

facilitates the communication between the different

components of the ITS and provides personalized

learning to the individual students. The operational

procedure of the multi-agent system has been

described and the overall functions of its

fundamental components have been illustrated. The

prototype provides dynamic curriculum sequencing

in a bottom up fashion using direct information

about the student preferences or learning styles and

relative information about the student learning

process as part of a group.

ACKNOWLEDGEMENTS

This research has been co-financed by the European

Union (European Social Fund – ESF) and Greek

national funds through the Operational Program

"Education and Lifelong Learning" of the National

Strategic Reference Framework (NSRF) (Funding

Program: “HOU”).

REFERENCES

Acampora, G., Loia, V., Gaeta, M., 2010. Exploring e-

Learning Knowledge Through Ontological Memetic

Agents. IEEE Computational Intelligence Magazine,

vol.5, no.2, pp. 66-77.

Ali, A, P., Dehghan, H., Gholampour, J., 2010. An Agent

Based Multilayered Architecture for E-learning

system. In Proceeding of E-Learning and E-Teaching

(ICELET). Second International Conference on IEEE,

pp.22-26.

Bellifemine, F., Caire, G., Pogg, A., Rimassa, G., 2003.

Jade-A White Paper. In EXP in search of innovation,

vol. 3, no. 3, 2003, pp. 6-19.

Bokhari, M., Ahmad, S., 2014. Multi-agent Based E-

Learning Systems: A Comparative Study. In

Proceedings of the 2014 International Conference on

Information and Communication Technology for

Competitive Strategies.

Bologna Working Group, 2005 A Framework for

Qualifications of the European Higher Education

Area.

Bishop B., Kiryakov, A., Ognyanoff, D., Peikov, I.,

Tashev, Z., Velkov, R., 2011. OWLIM: A family of

scalable semantic repositories. In Semant. web, vol. 2,

no. 1, pp. 33-42.

Dung, Q, P., Florea, M, A., 2011. An Architecture and a

Domain Ontology for Personalized Multi-Agent e-

Learning Systems. In Proceedings of the 2011 Third

International Conference on Knowledge and Systems

Engineering, pp. 181-185.

Foundation for Intelligent Physical Agents (FIPA). FIPA

ACL Message Structure Specification (2002), viewed

27 May 2014, http://www.fipa.org/specs/fipa00061/.

Hammami, S., Mathkour, H., Al-Mosallam, E.A., 2009. A

multi-agent architecture for adaptive E-learning

systems using a blackboard agent. 2nd IEEE

International Conference on Computer Science and

Information Technology, pp. 184-188.

Kalou, A., Solomou, G., Pierrakeas, C., Kameas, A., 2012.

An Ontology Model for Building, Classifying and

Using Learning Outcomes. 12th IEEE International

Conference on Advanced Learning Technologies, pp.

61-65.

Krathwohl, D., Anderson, L., (eds), 2001. A Taxonomy for

Learning, Teaching, and Assessing: A Revision of

Bloom's Taxonomy of Educational Objectives,

Abridged Edition.

LTSC Learner Model Working Group of the IEEE 2000.

Draft Standard for Learning Technology - Public and

Private Information (PAPI) for Learners (PAPI

Learner), IEEE p1484.2/d7, 2000-11-28.

McGuinness, D.L, van Harmelen, F., 2004. OWL Web

Ontology Language Overview, W3C Recommen-

dation, http://www.w3.org/TR/owl-features/

Nikolopoulos, G., Solomou, G., Pierrakeas, C., Kameas,

A., 2013. An Instructional Design Methodology for

Building Distance Learning Courses. 7th International

Conference in Open and Distance.

Nikolopoulos, G., Kalou, A., Pierrakeas, C., Kameas, A.,

2012a. Creating a LO Metadata Profile for Distance

Learning: An Ontological Approach. 6th Research

Conference, pp. 37-48.

Nikolopoulos, G., Solomou, G., Pierrakeas, C., Kameas,

A., 2012b. Modeling the Characteristics of a Learning

Object for Use within e-Learning Applications.

Proceedings of the Fifth Balkan Conference in

Informatics, pp. 112-117.

Nkambou, R., Mizoguchi, R., Bourdeau, J., (eds), 2010.

Advances in intelligent tutoring systems. Heidelberg:

Springer, 2010.

Noy N., McGuiness D., 2001. Ontology Development

101: A Guide to Creating Your First Ontology.

Stanford Knowledge Systems Laboratory Technical

Report KSL-01-05 and Stanford Medical Informatics

Technical Report SMi-2001-0880.

Panagiotopoulos, I., Kalou, A., Pierrakeas, C., Kameas,

A., 2012. An Ontological Approach for Domain

Knowledge Modeling and Management in E-Learning

Systems. In 1st AI in Education Workshop:

Innovations and Applications, pp. 95-104.

Panagiotopoulos I., Kalou, A., Pierrakeas, C., Kameas, A.,

2013. Adult student modeling for intelligent distance

learning systems. In Special Issue on AIAI 2012 of the

International Journal of Engineering Intelligent

CombiningAgentsandOntologiesforBuildinganIntelligentTutoringSystem

Systems for Electrical Engineering and

Communications, vol. 21, nos. 2/3.

Peña, A., Sossa, H., 2010. Semantic representation and

management of student models: An approach to adapt

lecture sequencing to enhance learning. In

Proceedings of the 9th Mexican international

conference on Advances in artificial intelligence: Part

I, pp. 175-186.

Polson, M., Richardson, J., 1988. Foundations of

Intelligent Tutoring Systems. Psychology Press.

Popham, W.J., 2011. Transformative Assessment in

Action: An Inside Look at Applying the Process.

ASCD, USA.

Smythe, C., Tansey, F., Robson, R., 2001. IMS Learner

Information Package Information Model Specification,

http://www.imsglobal.org/profiles/lipinfo01.html.

Yaghmaie, M., Bahreininejad, A., 2011. A context-aware

adaptive learning system using agents. In Expert

Systems with Applications, Vol. 38, Issue 4, pp. 3280-

3286.

CSEDU2015-7thInternationalConferenceonComputerSupportedEducation